CLAM: Safeguarding Authenticity and Addressing Implications for the Music Industry

The rapid rise of end-to-end AI music generation has introduced escalating challenges in detecting deepfake artist voice cloning, synthetic lyrics, and AI-altered acoustics, raising critical concerns around copyright, authenticity, and the preservation of human artistic expression. While early benchmarks like SONICS and models such as SpecTTTra offered foundational progress in identifying AI-generated songs, their reliance on a small number of generation sources limits generalization to newer models. In practice, we observe significant performance drops when applying existing detectors to out-of-distribution content from emerging generators such as Riffusion and Yue. To address this, we introduce Melody or Machine (MoM), a comprehensive benchmark dataset of 130,435 songs (6665.13 hours) synthesized using a diverse set of models and pipelines with controlled variations. MoM is curated to promote the development of robust, generalizable detection systems. Alongside this dataset, we introduce CLAM (Contrastive Learning for Audio Matching), a novel detection model that combines two pre-trained encoders: MERT for extracting expressive musical features and Wave2Vec2 for capturing vocal nuances. Their embeddings are fused through a learnable cross-aggregation module that extracts the learnable features from the intermediate layers. CLAM is trained using a combination of binary cross-entropy loss for detection and a triplet loss to align coherent music-vocal pairs in embedding space. CLAM achieves a state-of-the-art F1 score of 0.993 on the SONICS benchmark and 0.925 on our internal MoM dataset, setting a new state of the art in synthetic music forensics.